Article Preview
Top1. Introduction
The walking pattern of a person is termed as “Gait”. When any person walks, it can be noticed that the oscillation of swinging limbs repeat periodically. This periodical oscillation of swinging limbs varies from person to person. Therefore, this behaviour can be used for identification of a person and it is termed as “Gait Recognition”. Gait Recognition can be advantageous over the use of conventional physiological biometric traits:
On the other hand, physiological biometric traits need support from the person along with a complete monitored environmental setup for efficient and accurate recognition. Hence, physiological biometric traits alone are not very helpful, especially in surveillance systems. As Gait recognition system is capable of identifying a person from afar, it has become a hot topic for research in surveillance systems (Do et al., 2020; C. Benedek et al., 2018; H. Iwama et al., 2013). Surveillance monitoring system in public domain has huge video recordings and these recordings contain multiple angle views of a person. Variations in viewing angle, shooting venue, illumination, resolution ratio, frame ratio, walking speed, carrying bag or any object and type of clothes create a challenge in gait recognition (S. Sarkar et al., 2005; K. Bashir et al., 2010a). Out of these, variations in viewing angle is the most challenging as camera can only take pictures with one view angle, so the pictures captured by camera with cross-view look very different for the same subject (T. Connie et al., 2017; Goffredo M et al.,2010; W. Kusakunniran et al., 2012). Most of the time, video input to gait recognition system does not commensurate with the viewing angle of the enquirer and due to that performance of gait recognition system drops drastically (Yu S. et al., 2006). In order to handle these challenges, we propose a robust, efficient and lightweight system for gait recognition as shown in Fig. 1. We summarize this paper’s contributions as follows:
- •
We present two stage networks where the initial stage extracts optical flow feature of leg movement and linear regression is applied on it to achieve the state of the art results in view classification.
- •
The later stage uses gait energy image (GEI) and principle component analysis (PCA) combination for feature selection. Further, multilayer perceptron (MLP) is used to train the model for every angle separately.
- •
We have performed various experiments by considering different scenarios like normal walk, carrying bag, and variation in clothes in context with viewing angles.
Figure 1. A lightweight system for gait recognition
The remaining sections of this paper are organized as follows. The next section deals with related work in the gait recognition field. Section 3 describes the proposed method. Section 4 contains experimental results and analysis, and last section concludes this paper.